83 KiB
Google Maps Traffic/ETA and Mobility Data Products: How They Work, What Is Verified, and Privacy-Preserving Alternatives
Research memo for the sensing-layer paper
Prepared: 2026-11-14
Scope: Google Maps traffic and ETA; terminology; privacy claims and limits; commercial and public alternatives; implications for a first-party, privacy-first probe-data architecture.
Evidence labels used throughout
| Label | Meaning |
|---|---|
| Verified | Stated in a primary source (Google documentation, standard, peer-reviewed paper) and consistent across sources. |
| Company claim | Published by the vendor but not independently audited or reproducible from public data. |
| Not documented | No public evidence found in the sources reviewed. Absence of evidence, not proof of absence. |
| Misconception | Contradicted by the available primary sources, or a category error. |
One-sentence thesis. Google Maps is best understood as an aggregated probe-data and map-matching system that fuses device-side positioning with historical baselines and machine-learned ETA prediction; its public documentation does not describe it as a system that triangulates every handset from cell towers, and its public privacy story is stronger on data minimisation and retention than on formal, auditable guarantees for the traffic layer.
Key findings
- As publicly described, Google Maps traffic is a probe-data product, not a triangulation product. Live traffic comes from aggregate location data from people navigating; historical patterns and a graph neural network produce ETAs [Google-Maps101; Google-NavData; Google-ETA-GNN].
- Positioning is fused on the device. GPS, Wi-Fi, cell, and sensors are combined; GLA builds crowdsourced Wi-Fi/cell maps with a rotating device ID [Google-LocationData; Google-GeolocationAPI].
- Timeline is separate and off by default. It is now on-device with an optional encrypted backup and a three-month default auto-delete; live traffic does not require it [Google-Timeline2023; Google-TimelinePrivacy; Google-BusyAreas].
- The strongest privacy claim is narrow. Google claims differential privacy for busyness/popular times; no reviewed document claims DP for the traffic layer, and no ε is published [Google-BusyAreas].
- "Anonymized" is not a formal guarantee. Aggregation leaks membership; k-anonymity has well-known limits; membership inference on aggregate location data is feasible [Pyrgelis-2018; Sweeney-2002; deMontjoye-2013].
- Alternatives are complementary, not drop-in ground truth. HERE, TomTom, INRIX, Waze, Apple, telco, Placer/SafeGraph/Unacast, smart cards, and fleet data all have coverage and bias limits [HERE-RTT; TomTom-Prediction; SafeGraph-Bias; Vodafone-Cambridge].
- Penetration is the key unknown. Published work suggests 2–3% probe penetration can be enough for velocity, 3.5% for 5-minute flow, and 0.2% for hourly flow — but penetration is non-uniform and usually unknown to the data user [Herrera-2009; Seo-2015; Luo-2026].
- For our architecture: opt-in, on-device aggregation, local DP, secure aggregation, a k-threshold, no raw trajectories, and bounded retention. Claim "privacy-enhanced", not "anonymous" [Prio-2017; DAP-2026; RAPPOR-2014].
1. How it works
1.1 The stack in one page
Device-side positioning
GPS (satellite ranging) ─┐
Wi-Fi (AP scan / RTT) ─┼─► fused location estimate (OS / GLA)
Cell (cell-ID / timing) ─┤
Sensors (IMU, barometer)─┘
│
▼
Probe sample while navigating
(lat, lon, speed, heading, accuracy, timestamp, route context)
│
▼
Map matching to road segments
(snap noisy points to the road graph; reject outliers)
│
▼
Aggregation over users, segments, time bins, direction
(robust speed/travel-time statistics; sample counts)
│
├──► live traffic layer (green / orange / red)
├──► historical baseline (road × time-of-day × day-of-week)
└──► ML ETA model (live + historical + incidents + road attributes)
Google's public description is consistent across its consumer blog, help pages, and developer documentation: live traffic comes from aggregate location data from people navigating with Google Maps, and ETA prediction combines that live signal with historical traffic patterns and machine learning [Google-Maps101; Google-NavData; Google-RMI]. The exact sampling rate, penetration, map-matching algorithm, aggregation threshold, and per-road coverage are not public.
1.2 Positioning: GPS, Wi-Fi, cell, and sensors
Google's location stack does not rely on a single technology. Google states that Android estimates location using GPS, sensors (accelerometer, gyroscope, magnetometer, barometer), mobile network signals, and Wi-Fi signals, fused to produce the most accurate estimate available to apps [Google-LocationData]. Google Maps Help gives user-facing accuracy expectations: GPS is described as accurate to about 20 metres, Wi-Fi helps where GPS is weak (especially indoors), and cell towers can be accurate only to a few thousand metres [Google-LocationAccuracy].
The important architectural point is that these are device-side or OS-side positioning inputs. In the public model, Maps receives a fused location estimate; there is no public description of Maps reconstructing a handset's position from raw tower angles. The Google Geolocation API makes the fallback explicit: it accepts cell-tower and Wi-Fi access-point observations (including Wi-Fi MAC address and signal strength) and returns latitude, longitude, and an accuracy radius [Google-GeolocationAPI]. That is a server-side positioning service for devices without built-in geolocation; it is not evidence that Maps traffic is computed by tower triangulation.
Two distinctions matter for the paper:
- GPS is trilateration, not triangulation. A GPS receiver measures signal travel time to several satellites, converts time to range, and solves for position; four satellites are needed to resolve three spatial coordinates plus receiver clock error [GPSgov-Trilateration; NOAA-Geodesy]. The common phrase "GPS triangulation" is technically wrong.
- Wi-Fi positioning is usually fingerprinting or multilateration, not triangulation. Ordinary Wi-Fi positioning matches the observed set of access-point identifiers and signal strengths against a database of known access points. Wi-Fi RTT (round-trip time) measures distances to three or more access points and uses multilateration, typically accurate to 1–2 metres under supported conditions [Android-RTT]. Neither is angle-based triangulation.
Google also operates Google Location Accuracy (GLA), formerly Google Location Services. When GLA is on, Android periodically collects GPS and information about Wi-Fi access points, mobile networks, and device sensors using a temporary rotating device identifier that is not associated with a specific person. Google says it uses this to improve location accuracy and to build crowdsourced maps of Wi-Fi access points and mobile network towers [Google-LocationData]. This is one of the clearest public descriptions of the "Google has a database of Wi-Fi and cell locations" claim. It is still not the same as passive sniffing of nearby people's device MAC addresses: the object being mapped is the infrastructure (access points, towers), and the identifier used for the device's own reports rotates.
1.3 Probe data and map matching
A probe is a position/speed report from a participating device or vehicle. In traffic engineering, the same idea is called floating car data (FCD): vehicles act as moving sensors, and their trajectories are map-matched to the road network to estimate link speeds and travel times [TomTom-API; HERE-Probe]. Google's consumer description uses the phrase "aggregate location data" rather than "floating car data", but the logic is the same.
Google's navigation-data help page is unusually specific about what is collected during navigation [Google-NavData]:
- GPS location;
- navigation details, such as the route taken;
- sensor data from the device, such as the barometer.
Collection starts shortly after the user taps Start (or begins moving with glanceable directions on) and stops after arrival. Google says it uses this data "by itself or combined with others" for navigation, alternate routes, and real-time updates such as traffic, disruptions, and weather [Google-NavData].
The public description stops short of the pipeline details that matter for reproducibility:
- How often samples are taken (the 2009 launch described anonymous speed updates; current docs do not give a cadence).
- How many probes are needed before a road segment is published.
- How probes are filtered for outliers, stops, mode (driving vs cycling vs transit), and GPS error.
- How map matching handles parallel roads, tunnels, parking garages, and urban canyons.
- Whether probes from Android Auto, iOS, Waze, or third-party Maps Platform customers are pooled, and with what weighting.
These are not documented in the sources reviewed. That absence is itself a finding: Google publishes the product behaviour, not the measurement methodology.
1.4 Live traffic, historical traffic, and ETA
Google separates three related products:
- Live traffic. Current conditions on road segments, rendered as green ("no traffic delays"), orange ("medium amount of traffic"), red ("traffic delays"), and darker red for slower speeds [Google-TrafficLegend]. Google's 2020 explanation says this is powered by aggregate location data from people navigating [Google-Maps101].
- Historical / typical traffic. Per-road patterns by time of day and day of week. Google says it combines a historical traffic database with live conditions using machine learning, and that since COVID-19 it prioritises patterns from the last two to four weeks while deprioritising older patterns [Google-Maps101].
- ETA prediction. The estimated travel time for a route, combining live conditions, historical patterns, road attributes (paved/unpaved, size, directness), and incidents. Google's 2020 blog claims ETA predictions are "consistently accurate for over 97% of trips" and that a DeepMind partnership reduced inaccurate ETAs by up to 50% in some cities [Google-Maps101; DeepMind-2020]. The peer-reviewed CIKM paper describes a graph neural network over "supersegments" (connected road subgraphs) deployed in production, with significant reductions in negative ETA outcomes (40%+ in Sydney) [Google-ETA-GNN].
Google's developer documentation adds a useful rule of thumb: short routes depend largely on current real-time information; longer routes use AI modelling where nearby segments are predicted from real-time data and more distant segments rely more heavily on historical patterns; roads with limited real-time signals rely more heavily on historical data [Google-RMI]. That is a sensible engineering statement, and it implies that "live traffic" coverage is not uniform: some roads are data-rich, others are model-filled.
Incidents come from multiple sources: user reports in Maps, Waze reports (Waze says its alerts reach both Waze and Google Maps drivers), and authoritative feeds from road operators [Google-Maps101; Waze-Cities; Google-RMI]. This fusion is why Google Maps can show a crash or closure even when probe speeds alone would be ambiguous.
1.5 Timeline, Popular Times, and Busy Areas
Google's Timeline (formerly Location History) is a separate, opt-in product. It is off by default, and when enabled it saves precise device location regularly — including when Google apps are not in use — to build a personal map of visits and routes [Google-TimelinePrivacy; Google-TimelineManage]. Since the December 2023 announcement, Timeline is designed to be saved on the device, with an optional end-to-end encrypted cloud backup; the default auto-delete period was reduced to three months, with 18- and 36-month options still available [Google-Timeline2023; Google-TimelineManage].
Popular Times, wait times, and visit duration are computed from aggregated and anonymized data from users who have opted in to Timeline. Google says the data is shown only when there are enough visits, and it explicitly claims differential privacy for busyness data; it also says it does not calculate busyness for residential places and suppresses busyness when it lacks enough data [Google-BusyAreas]. The same help page describes Busy Areas as an aggregate of live busyness trends from places in a predefined area of interest, never the exact location of an individual [Google-BusyAreas].
There is a documented tension here. Google's location-data page says Timeline data can be used to "show community trends, such as popular times and transit crowdedness, based on aggregated location data" [Google-LocationData], while the Timeline help page says Timeline is now saved on the device [Google-TimelinePrivacy; Google-TimelineManage]. Google does not publicly specify how on-device Timeline data and aggregated community trends are reconciled, or what leaves the device and under which settings. That is a legitimate question for the paper, but it should be framed as an ambiguity in public documentation, not as proof that raw trajectories are uploaded.
Also note the separation: Google's live-traffic description relies on navigation data, not on Timeline. Timeline is used for busyness and personal history; it is not required for the traffic layer [Google-NavData; Google-BusyAreas]. This matters because critics often conflate "Location History" with "traffic data collection". They are different settings with different defaults.
1.6 What Google actually does, step by step
A defensible summary for the paper:
- The device/OS fuses GPS, Wi-Fi, cell, and sensor signals into a location estimate with an accuracy radius [Google-LocationData; Google-GeolocationAPI].
- While the user navigates, Maps collects GPS, route, and sensor data and stops at arrival [Google-NavData].
- Google map-matches probe samples to road segments and aggregates speeds/travel times across users and time bins [Google-Maps101; Google-RMI].
- Live segment speeds feed the traffic layer; historical patterns are learned per road and time period; the two are combined by ML for ETA [Google-Maps101; DeepMind-2020; Google-ETA-GNN].
- Incidents from users, Waze, and road operators are fused into routing and ETA [Google-Maps101; Waze-Cities; Google-RMI].
- Separately, opt-in Timeline data supports personal history and, in aggregate, Popular Times/Busy Areas with a DP claim [Google-TimelinePrivacy; Google-BusyAreas].
That is the publicly described model. Everything below the line — sample counts, thresholds, weighting, error models, DP parameters — is either proprietary or unpublished.
2. Terminology: triangulation, trilateration, multilateration, fingerprinting, probe data
2.1 Definitions
| Term | What is measured | How position is obtained | Typical example | Common misuse |
|---|---|---|---|---|
| Triangulation | Angles (bearings) from known points | Intersect bearing lines | Surveying with a theodolite; classical geodetic networks | Calling any multi-signal positioning "triangulation" |
| Trilateration | Distances (ranges) from known points | Intersect distance circles/spheres | GPS/GNSS; some terrestrial ranging systems | Calling GPS "triangulation" |
| Multilateration | Multiple range or range-difference measurements | Solve for the point that best fits all measurements | Wi-Fi RTT with 3+ APs; TDOA systems; some cellular positioning | Confusing it with triangulation or with fingerprinting |
| Fingerprinting | A radio "signature" (e.g., set of Wi-Fi BSSIDs and RSSI values) | Match the observed signature to a pre-built database of known locations | Google Geolocation API; Wi-Fi positioning | Treating it as geometric triangulation |
| Cell-ID positioning | Serving cell / tower identity | Use the known or estimated location of the cell | Coarse mobile location | Assuming tower-level accuracy is metres |
| Probe data | Positions, speeds, headings, timestamps from participating devices/vehicles | Map-match and aggregate; not a positioning method itself | Floating car data; Google Maps navigation data; HERE/TomTom probes | Calling probe data "triangulation" |
| Map matching | A noisy point sequence plus a road graph | Snap points to the most plausible road segment/path | Traffic speed estimation; routing | Assuming every probe is correctly matched |
| Aggregation | Many probe records | Combine into segment × time-bin statistics | Traffic layer; heatmaps; demand indices | Assuming aggregation is a formal privacy guarantee |
Sources for the geometric definitions: NOAA's geodetic tutorial defines triangulation as measuring angles in a chain of triangles, and trilateration as measuring the lengths of triangle sides [NOAA-Geodesy]. GPS.gov describes GPS positioning as trilateration from satellite ranges [GPSgov-Trilateration]. Android's Wi-Fi RTT documentation describes multilateration from distances to three or more access points [Android-RTT]. Google's Geolocation API documentation describes cell-tower and Wi-Fi access-point inputs and an accuracy radius, which is consistent with fingerprinting/weighted positioning rather than angle-based triangulation [Google-GeolocationAPI].
2.2 What Google actually does vs the common misconception
| Misconception | What the sources actually support |
|---|---|
| "Google Maps triangulates every phone from cell towers." | Google fuses GPS, Wi-Fi, cell, and sensors on the device, and uses aggregate location data from people navigating for traffic [Google-LocationData; Google-Maps101]. No reviewed Google document describes a tower-triangulation pipeline for Maps traffic. |
| "Google Maps traffic requires Location History." | Live traffic is described as coming from navigation data; Timeline is a separate opt-in setting used for personal history and busyness [Google-NavData; Google-BusyAreas]. |
| "Incognito mode stops Google from using you for traffic." | Google's Incognito documentation says it stops saving search/navigation history, updating Timeline, and personalizing Maps; it does not state that anonymous navigation/traffic measurement stops [Google-Incognito]. |
| "Google uses Bluetooth MAC sniffing to count cars." | No reviewed Google documentation describes passive Bluetooth or client-MAC sniffing as a Maps traffic source. Google's documented radio positioning uses Wi-Fi access-point observations and cell-tower data [Google-GeolocationAPI; Google-LocationData]. That is infrastructure-side fingerprinting, not passive counting of nearby people's device MACs. |
| "Anonymized means differentially private." | Google uses generalization/k-anonymity and DP as distinct techniques in its general anonymization page [Google-Anonymization]. DP is explicitly claimed for busyness data [Google-BusyAreas], but no reviewed document claims DP for the traffic layer. |
| "More probes always mean better traffic." | Probe penetration is spatially and temporally non-uniform; accuracy depends on sample size, map matching, and road coverage, not just volume [Luo-2026; Seo-2015; Mazare-2011]. |
The most important correction for our paper: as publicly described, Google's traffic product is a probe-data product, not a triangulation product. The radio-positioning layer exists, but it is an input to the device's location estimate. The traffic layer is built from map-matched movement observations and historical models. If we criticise "triangulation" as a sensing layer, we should be precise about which layer we mean.
2.3 Recommended terminology for our paper
- Use "aggregated probe data with map matching" for our proposed sensing layer. This describes what we actually collect and how we turn it into road/area statistics.
- Use "multi-receiver multilateration" only if we estimate positions from ranges or range differences measured by multiple receivers.
- Use "fingerprinting" if we match observed Wi-Fi/Bluetooth signatures to known locations.
- Use "triangulation" only when angles are measured. If we use it colloquially, define it explicitly and note that GPS is trilateration.
- Avoid "anonymous" unless we have a formal guarantee and an audit. Prefer "privacy-enhanced", "aggregated with a k-threshold", or "local differential privacy with a stated ε".
- Avoid "counts people". Use "estimated devices" and state that device penetration is unknown. Multiple devices per person and shared devices are both real.
A useful sentence for the paper:
Google's public documentation does not describe traffic as computed by triangulating handsets from cell towers; it describes fusing device-side location signals and aggregating map-matched probe trajectories from people who navigate with the app. The relevant privacy question is therefore not "can Google locate a handset?" but "what is retained, at what granularity, for how long, and under which formal guarantee?"
3. What is verified vs claimed
3.1 Claim-by-claim assessment
| # | Claim | Status | Evidence | Caveat |
|---|---|---|---|---|
| 1 | Google Maps live traffic uses aggregate location data from people navigating | Verified | Google's 2020 Maps 101 post and navigation-data help page [Google-Maps101; Google-NavData] | Google does not publish sample counts, penetration, or per-road coverage. |
| 2 | Google combines live traffic with historical patterns using ML | Verified | Google Maps 101; DeepMind blog; ETA GNN paper [Google-Maps101; DeepMind-2020; Google-ETA-GNN] | The relative weight of live vs historical is not public. |
| 3 | Google uses a graph neural network for ETA | Verified | CIKM 2021 paper describes production deployment [Google-ETA-GNN] | The paper reports aggregate improvements, not per-city production metrics. |
| 4 | Google's ETA predictions are accurate for "over 97% of trips" | Company claim | Google Maps 101 [Google-Maps101] | "Accurate" is undefined; no independent audit or error distribution is published. |
| 5 | DeepMind reduced inaccurate ETAs by up to 50% in some cities | Company claim | Google Maps 101; DeepMind blog [Google-Maps101; DeepMind-2020] | City-specific, not a global guarantee; the baseline and metric are not fully specified. |
| 6 | Google collects GPS, route, and sensor data during navigation | Verified | Navigation-data help page [Google-NavData] | Collection starts at Start and stops at arrival; the page does not give a sampling cadence. |
| 7 | Google uses Wi-Fi and cell signals for location | Verified | Location-data page; Geolocation API docs [Google-LocationData; Google-GeolocationAPI] | This is positioning, not proof that traffic uses raw Wi-Fi/cell observations. |
| 8 | Google builds crowdsourced maps of Wi-Fi APs and cell towers with a rotating device ID | Verified | Location-data page [Google-LocationData] | The object mapped is infrastructure; the device identifier rotates. |
| 9 | Google uses differential privacy for Maps busyness data | Verified as a Google statement | Busy Areas help page [Google-BusyAreas] | No ε, sensitivity, composition, or independent audit is published. |
| 10 | Google uses differential privacy for the traffic layer | Not documented | No reviewed Google document claims this | Absence of evidence; do not assert either way without a source. |
| 11 | Google describes traffic data as aggregated/anonymized | Company claim | Google uses "aggregated and anonymized" language [Google-Maps101; Google-BusyAreas] | "Anonymized" is not a formal privacy model; aggregation can leak membership [Pyrgelis-2018]. |
| 12 | Google deletes navigation data after processing | Company claim | Navigation-data help page [Google-NavData] | No timeframe, audit, or backup-retention detail is given for this specific data. |
| 13 | Google permanently deletes trip start/end points | Historical company claim | 2009 launch post and contemporaneous coverage [Google-2009; Google-AU-2009] | A 2009 statement; current docs say "after processing, it's deleted" without a timeframe. |
| 14 | Timeline is on-device and off by default | Verified | 2023 blog; Timeline help pages [Google-Timeline2023; Google-TimelinePrivacy; Google-TimelineManage] | Cloud backup is optional; other settings (Web & App Activity, ads) can still save location. |
| 15 | Incognito stops traffic contribution | Not documented | Incognito help page lists only account-history/personalization effects [Google-Incognito] | Do not claim it does or does not; say Google does not document it. |
| 16 | Google uses Bluetooth MAC sniffing for traffic | Misconception / not documented | No reviewed Google source describes it | Do not attribute MAC sniffing to Google Maps. |
| 17 | Google triangulates handsets from cell towers for traffic | Misconception | Public model is fused device positioning + aggregate probes [Google-LocationData; Google-Maps101] | Tower data is one positioning input, not the traffic algorithm. |
| 18 | Google Maps traffic coverage is global | Company claim, underspecified | Maps is described as covering 220+ countries; traffic-layer coverage is not published [Google-Maps101] | Country coverage of Maps ≠ traffic-layer coverage of every road. |
3.2 What Google does not disclose
The following are not documented in the public sources reviewed. They are the questions a rigorous paper should ask rather than answer by assumption:
- Penetration and sample size per road segment and time bin.
- Minimum probe threshold before a segment is published.
- Map-matching and outlier filters, including how stops, mode, and GPS jumps are handled.
- Whether Wi-Fi/cell observations contribute directly to traffic probes, or only to the device's location estimate.
- DP parameters (ε, δ, sensitivity, composition) for busyness, and whether traffic uses any formal DP mechanism.
- Retention period for navigation data beyond "after processing".
- Traffic-layer coverage by country, road class, and time of day.
- Independent audits of accuracy or privacy claims.
This is not a reason to accuse Google of wrongdoing. It is a reason to treat its published claims as company claims and to design our own system so that these parameters are explicit.
3.3 A note on the 2009 launch
Google's 2009 announcement is the origin of its "crowdsourced traffic" story. It described GPS-enabled phones with My Location enabled sending anonymous speed and location data; Google combined thousands of such reports into live traffic. Google also said it found the start and end points of every trip and permanently deleted that data so that even Google would not have access to it [Google-2009; Google-AU-2009]. That is a strong early privacy claim, and it is worth citing as historical context. But it is a 2009 company statement, not a current technical specification. Current documentation uses softer language ("after processing, it's deleted") and does not publish a retention period [Google-NavData].
4. Privacy model
4.1 What Google says it does
Google's main Privacy Policy is the umbrella document; the location-data page provides Maps-specific detail [Google-PrivacyPolicy; Google-LocationData].
| Mechanism | Where Google states it | What it is meant to protect | What it does not establish |
|---|---|---|---|
| Aggregation | Traffic and busyness descriptions [Google-Maps101; Google-BusyAreas] | Individual probe is hidden in a group | Membership inference, small-cell leakage, or bias |
| Anonymization / pseudonymization | Location-data and anonymization pages [Google-LocationData; Google-Anonymization] | Removes or generalizes identifiers | A formal guarantee; "anonymous" is undefined |
| Generalization → k-anonymity, l-diversity | Anonymization page [Google-Anonymization] | Re-identification via quasi-identifiers is harder | Homogeneity, background knowledge, composition attacks |
| Differential privacy | Busy Areas help page [Google-BusyAreas] | Membership inference with a mathematical guarantee | Utility loss; no ε published; scope limited to busyness |
| Rotating identifiers | GLA description [Google-LocationData] | Limits long-term linkability of device reports | Linkability within a session; other identifiers |
| On-device Timeline | 2023 blog; Timeline help [Google-Timeline2023; Google-TimelinePrivacy] | Keeps personal history off Google servers by default | Backups, other settings, device compromise |
| Retention / auto-delete | Timeline and retention pages [Google-TimelinePrivacy; Google-Retention] | Limits long-term accumulation | Data already used; backups; derived aggregates |
| Incognito mode | Maps help [Google-Incognito] | Stops account history and personalization | Does not document stopping anonymous traffic measurement |
| Location data collection toggle | iOS Maps help [Google-LocationToggle] | Allows opting out of anonymous location collection | Platform-level location services may still operate |
One nuance: Google's navigation-data page says it does not associate map updates with the user's Google Account or device, but it also lists data that may be account-associated — direction searches, ads shown during navigation, and Gemini activity [Google-NavData]. The privacy boundary is therefore per data type, not per user. "Navigation data is not associated with your account" does not mean "no location data is associated with your account."
4.2 Aggregation, k-anonymity, and DP are not the same thing
This is the core conceptual distinction for the paper.
- Aggregation combines many records. It is a data-processing step, not a privacy guarantee. If a cell has one dominant contributor, or if an adversary can compare releases with and without a target, aggregation can leak membership. Pyrgelis et al. show that membership inference on aggregate location time-series is feasible, and that some defences fail or impose large utility costs [Pyrgelis-2018]. De Montjoye et al. show that four random spatio-temporal points can uniquely characterise 95% of individuals in a coarse mobility dataset [deMontjoye-2013]. Aggregation raises the cost of re-identification; it does not make it impossible.
- k-anonymity requires each released record to be indistinguishable from at least k−1 others on the quasi-identifiers [Sweeney-2002]. It protects against direct re-identification only if the quasi-identifier set and the adversary's background knowledge are modelled correctly. It does not protect against attribute disclosure, homogeneity, or differencing attacks. Google's general anonymization page describes k-anonymity and l-diversity as techniques it uses [Google-Anonymization]; Google's search location handling uses a concrete spatial threshold — an area of at least 3 km² or 1,000 people — to avoid revealing a precise search location [Google-LocationData]. That is a useful precedent for a spatial k-threshold, but it is not a statement about traffic data.
- Differential privacy (DP) requires the output distribution to change by at most a bounded factor when one individual's data is added or removed [Dwork-2006; Dwork-Roth-2014]. DP provides a formal membership guarantee and composes across releases, but it requires calibrated noise, a stated ε/δ, and a privacy budget. Local DP randomizes on the device before transmission, so the server never sees the raw value [RAPPOR-2014]. Central DP assumes a trusted curator. Secure aggregation is complementary: it hides individual contributions cryptographically, but does not add noise, so it does not by itself provide DP [Prio-2017; Bonawitz-2017; DAP-2026].
Google's public claims mix these categories. The busyness page says "differential privacy", which is the strongest claim [Google-BusyAreas]. The traffic pages say "aggregated" and "anonymized", which are weaker and undefined [Google-Maps101; Google-NavData]. A paper that treats "Google says it anonymizes traffic" as equivalent to "Google has a DP guarantee for traffic" would overstate the evidence.
4.3 Retention: what is actually documented
| Data | Documented retention | Source |
|---|---|---|
| Timeline (Location History) | Auto-delete after 3, 18, or 36 months; default reduced to 3 months in the 2023 update | [Google-TimelinePrivacy; Google-TimelineManage; Google-Timeline2023] |
| Web & App Activity location | 30 days for IP-based and device location | [Google-LocationData; Google-Retention] |
| Navigation data | "After processing, it's deleted" — no timeframe published | [Google-NavData] |
| Deleted account data | Deletion generally takes ~2 months; backups may retain for up to 6 months | [Google-Retention] |
| Advertising server logs | Partial IP removed after 9 months; cookie information after 18 months | [Google-Retention] |
| TfL Oyster journey history (comparison) | 8–9 weeks, then disassociated from the card | [TfL-Access] |
| TfL contactless journey history (comparison) | 13 months | [TfL-Access] |
The retention picture is mixed. Google has improved consumer controls for Timeline (on-device, shorter default), but the traffic/navigation pipeline is still described in vague terms. For a privacy-first architecture, "deleted after processing" is not a specification; we should be able to state exactly when raw data is destroyed, what derived aggregates remain, and how long they are retained.
4.4 Where the privacy story is weaker than the marketing
- "Anonymized" is doing a lot of work. Google uses the word for traffic and busyness, but does not publish the de-identification method, the quasi-identifier set, the k threshold, or a re-identification test. Under GDPR-style analysis, pseudonymized data remains personal data; anonymization is a high bar [ISO-20889; NIST-800-188].
- In the Maps material reviewed, DP is claimed only for busyness. That is a meaningful claim, but it is narrow. There is no public ε, no sensitivity analysis, and no public independent audit. A DP claim without parameters is not reproducible.
- Aggregation is not a guarantee. The membership-inference literature shows that aggregate location time-series can leak whether a target contributed [Pyrgelis-2018]. Small cells, regular routines, and differencing across time are the attack surfaces.
- On-device Timeline is a real improvement, but the boundary is fuzzy. If Timeline data is on-device, how does it support aggregated community trends? Google's own pages say both things without reconciling them [Google-TimelinePrivacy; Google-LocationData]. The answer may be backups, separate settings, or aggregated on-device computation; Google does not say.
- Incognito is not a traffic opt-out. It stops account history and personalization, not necessarily anonymous measurement [Google-Incognito]. Users may reasonably assume otherwise; the documentation should be clearer.
- No public independent audit. Accuracy and privacy claims are self-reported. That is normal for commercial maps, but it means the paper should label them as claims.
4.5 A checklist for distinguishing marketing from verification
When a vendor says... ask:
| Phrase | Question that turns it into evidence |
|---|---|
| "Anonymous" | Which formal model? k-anonymity, DP, or just de-identified? What is k or ε? Has an independent party tested re-identification? |
| "Anonymized" | Is the data irreversibly de-identified, or pseudonymized? Can it be joined back to an identifier? |
| "Differential privacy" | Local or central? What ε/δ? What sensitivity? How is the budget composed across releases? |
| "On-device" | What leaves the device? Are there backups? Are aggregates uploaded? Under which settings? |
| "Opt-in" | Is it default-off? Is consent granular (traffic vs ads vs research)? Can it be withdrawn? |
| "Deleted" | Deleted from which systems, when, and from backups? Are derived aggregates retained? |
| "Aggregated" | What is the spatial/temporal bin? What is the minimum count? Are small cells suppressed? |
| "Real-time" | What is the update cadence, latency, and coverage? Is it live data or a model? |
These questions are also the specification for our own system.
5. Alternatives: other mobility-data products
5.1 Comparison table
| Source / product | What it measures | Primary data source | Coverage / resolution | Accuracy and bias notes | Documented privacy posture |
|---|---|---|---|---|---|
| Google Maps traffic/ETA | Segment speeds, travel times, ETAs, incidents | Aggregate location data from people navigating; GPS/Wi-Fi/cell/sensor positioning; historical patterns; DOT/Waze incidents [Google-Maps101; Google-NavData; Google-RMI] | Global consumer Maps; traffic-layer coverage not published; per-road segments | 97% ETA claim is company-reported; penetration unknown; urban/device bias | Aggregated/anonymized; DP claimed for busyness, not traffic; Timeline on-device; retention controls [Google-BusyAreas; Google-Timeline2023] |
| HERE Real-Time Traffic / Traffic Analytics | Flow speeds, congestion, incidents, historical probe statistics | Connected-car probes, vehicle sensors, fixed sensors, government feeds; 160+ auto-grade probe sources [HERE-RTT; HERE-OnePager; HERE-Probe] | 75 countries; flow updated every minute; incidents every 2 minutes; +4 h prediction [HERE-RTT; HERE-OnePager] | Multi-source fusion; confidence field distinguishes real-time vs historical vs speed-limit-only data [HERE-RTT]; fleet mix bias | Commercial; aggregated/anonymized probe data; customer contracts govern use |
| TomTom Traffic / Traffic Stats | Flow speeds, travel times, incidents, historical traffic stats, O/D | Floating car data from ~600 million probes: PNDs, in-dash systems, apps, fleets [TomTom-Prediction; TomTom-Stats; TomTom-API] | Global; country coverage varies by product; 5-min historical epochs; sample size and percentiles published in Traffic Stats [TomTom-API] | FCD is a non-probability sample; device/app mix bias; TomTom publishes sample counts and percentiles, which is good practice | Commercial; anonymized FCD; aggregated historical products |
| INRIX AI Traffic | Real-time speeds, travel times, incidents, historical analytics | Probe data, 20 years of history, "trillions of data points"; all road classes including non-TMC [INRIX] | Claims all roads, including rural and arterial; country coverage varies | "Most accurate" is a company claim; probe mix and penetration not public | Commercial; aggregated; privacy policy applies |
| Waze for Cities | Driver-reported incidents, closures, speeds, traffic flow | Crowdsourced GPS + driver reports; ~200 million drivers [Waze-Cities] | Strong in populated areas; weaker where few Wazers; incident-centric | Self-selected drivers; reporting bias; rural weakness noted by DOT practitioners [IowaDOT-2020] | Free data-sharing with cities; "anonymized driver speed and traffic flow data" [Waze-Cities]; Google-owned |
| Apple Maps | Traffic, routing, ETAs, incidents | iPhone probes; random identifiers; on-device personalization [Apple-Privacy] | Global consumer Maps; coverage not published | No public penetration or accuracy data; strong consumer privacy posture | Random identifiers that reset; location fuzzing to a less-exact area within 24 h; no history of searches/where you've been [Apple-Privacy] |
| Telco mobility data (Orange Flux Vision, Telefónica, Vodafone) | Population presence, flows, O/D matrices, dwell patterns | Mobile network signalling / cell-tower events [Telefonica; Orange-Flux; Vodafone-Cambridge] | National/regional; cell-tower granularity; rural areas coarser; low-density areas may be suppressed | Subscriber market-share bias; cell-level spatial error (hundreds of metres to kilometres) [Google-LocationAccuracy]; background events uneven; low-density suppression [Vodafone-Cambridge] | Aggregated/anonymized; GDPR/CNIL claims; Orange calls its process "irreversible anonymization" [Orange-Flux]; legal basis and thresholds vary |
| Placer.ai | Footfall, visits, trade areas, audience behavior | Panel of tens of millions of mobile devices, sourced from vetted app partners [Placer-Data] | US-focused; POI/venue level; daily refresh | Panel/app bias; extrapolation; company claims >90% correlation with ground truth [Placer-Data] | Strips MAIDs; k-anonymity of 50 (at least 50 devices); no user-level display; opt-in/opt-out required of partners [Placer-Data; Placer-Trust] |
| SafeGraph | POI visits, foot traffic, spend patterns | Licensed data from third-party aggregators; no SDK; opt-in device data [SafeGraph-Privacy; SafeGraph-Warren] | US visit panel; POI/census-block level | Small fraction of US population; documented spatial/temporal bias [SafeGraph-Bias]; extrapolation error | Aggregated; DP noise claimed in Patterns; sensitive-category suppression; no individual-level sale [SafeGraph-Warren] |
| Unacast | Footfall, visitor journeys, trade areas, demographics | Opt-in mobile app data; first-party aggregation; ML model over aggregates and contextual features [Unacast-Method; Unacast-Legal] | US-focused; POI/hexagon level; 48–72 h delay; not real-time [Unacast-Legal] | Panel bias; ML extrapolation; no real-time use | Aggregated; no real-time tracking; sensitive-place suppression (PrivacyCheck); opt-out honored [Unacast-Legal] |
| Transit smart cards (TfL Oyster/RODS) | Transit entries/exits, O/D flows, station counts | Fare-gate taps, Oyster/contactless, rolling O/D surveys [TfL-OpenData] | London; station/line level; 5% Oyster journey sample (2009); RODS ongoing [TfL-OpenData] | Fare-paying transit users only; no car traffic; tap-out is not universal, so some journeys are inferred/auto-completed [TfL-Oyster] | Pseudonymized journey data; Oyster history retained 8–9 weeks then disassociated; contactless 13 months [TfL-Access; TfL-Oyster] |
| GPS fleet telematics (Geotab/Altitude) | Vehicle movements, speeds, corridor volumes, commercial vehicle patterns | First-party telematics devices in commercial fleets; millions of vehicles [Geotab-FAQ; Geotab-Privacy] | Commercial vehicles; corridor/regional aggregates; primary-road coverage claims ~99.9% in some products [Geotab-FAQ] | Commercial fleet bias (trucks/vans, not passenger cars); route and vocation bias | Aggregated/de-identified; minimum-count thresholds; geospatial obfuscation; salted/hashed VINs; no individual vehicle data [Geotab-FAQ] |
| First-party probe data (our proposal) | Segment speeds, travel times, area footfall/dwell | Opt-in app/device probes; on-device binning; local DP; secure aggregation | Our deployment footprint; not city-wide unless calibrated | Penetration unknown; device ≠ person; needs ground-truth calibration | Design-dependent; can be strong if formal DP + k-threshold + no raw trajectories + retention limits |
5.2 Notes on coverage, accuracy, and bias
Coverage is not the same as accuracy. HERE, TomTom, and INRIX claim broad road coverage; Waze claims a large driver community; Google claims global Maps usage. None of these claims tells us the probe penetration on a specific road at a specific hour. Penetration is the key parameter for turning probe samples into population estimates, and it is usually unknown to the data user [Mazare-2011; Luo-2026].
All commercial probe datasets are non-probability samples. They come from people who use a particular app, drive a particular vehicle, or carry a particular device. That creates selection bias by age, income, smartphone platform, vehicle type, and geography. SafeGraph's panel has been shown to have systematic spatial and temporal bias [SafeGraph-Bias]. Placer and Unacast try to correct this with ML and ground-truth calibration, but the correction is proprietary and not independently reproducible [Placer-Data; Unacast-Method].
Telco data is population-scale but spatially coarse. It is attractive for O/D matrices and regional flows, not for lane-level traffic. Cell-tower granularity, subscriber market share, and low-density suppression limit its use for street-level sensing [Telefonica; Orange-Flux; Vodafone-Cambridge]. Orange's claim of "irreversible anonymization" is a legal/technical position, not an independently verified DP guarantee [Orange-Flux].
Smart cards are excellent for transit, not a substitute for road traffic. They provide high-quality O/D data for fare-paying transit users, but they cannot measure road traffic and they exclude non-fare users. TfL's 5% Oyster sample and RODS are good examples of aggregate release with explicit retention limits [TfL-OpenData; TfL-Access].
Fleet data is high-quality but commercially biased. Geotab's Altitude product comes from commercial vehicles, so it over-represents trucks, vans, and fleet routes. It is useful for freight and corridor planning, not for general passenger-car traffic [Geotab-FAQ; Geotab-Privacy].
Waze is incident-rich and coverage-skewed. It is strong where drivers use Waze and report incidents; it is weaker in rural or low-usage areas. DOT practitioners have noted this limitation [IowaDOT-2020]. It is also Google-owned, so it is not an independent alternative to Google.
Apple is the closest consumer-privacy comparator. Apple says Maps data is associated with random identifiers that reset, location is fuzzed within 24 hours, and Apple does not retain a history of searches or where you have been [Apple-Privacy]. Apple does not publish traffic coverage or penetration, so its accuracy claims are not directly comparable. It is a useful example of "privacy by architecture" language, but it is still a company claim.
5.3 Which source for which job?
| Job | Best-fit sources | Why |
|---|---|---|
| Live segment speeds / ETA | Google, HERE, TomTom, INRIX, Waze | Broad probe bases and mature map matching; use at least two for cross-checking |
| Historical traffic patterns | HERE, TomTom, INRIX, Google | Multi-year archives; TomTom/HERE publish sample counts/percentiles |
| Incident detection | Waze, Google, HERE, INRIX | Crowd reports + operator feeds; Waze is incident-centric |
| Regional O/D / population flows | Telco (Orange, Telefónica, Vodafone) | Population-scale signalling data; coarse spatial resolution |
| Retail footfall / trade areas | Placer, SafeGraph, Unacast | Panel-based visit estimates; calibrate with store counts |
| Transit O/D | TfL/agency smart-card data | High-quality mode-specific data; not general traffic |
| Freight / commercial vehicle flows | Geotab/Altitude, HERE, TomTom | Fleet telematics; commercial bias |
| Our own venues / corridors | First-party probe + ground counters | Control over consent, aggregation, retention, and calibration |
6. From aggregates to heatmaps and demand forecasts
6.1 Spatial and temporal binning
A probe dataset becomes a heatmap or demand index through four operations:
- Map matching (for road data): snap each point to a road segment and direction.
- Spatial binning: use road segments for traffic, or a hierarchical grid for area data. H3 is a common open-source hexagonal hierarchy; it supports multi-resolution aggregation and stable joins across datasets [H3]. Geohash or administrative boundaries are alternatives.
- Temporal binning: 1–5 minutes for real-time traffic, 15 minutes for operational dashboards, 1 hour for demand forecasts. The bin size must be large enough to satisfy the k-threshold and small enough to be useful.
- Statistic choice: count, median/mean speed, travel time, dwell time, or a normalized index. Use robust statistics (median, trimmed mean) and publish sample counts/uncertainty.
For footfall, the raw output is usually estimated devices per bin, not people. A device may be a phone, tablet, or vehicle; one person may carry multiple devices; a device may be shared. The paper should say "estimated devices" at least once and explain why.
6.2 Expansion and bias correction
There are three ways to turn a sample into a population estimate:
- Design-based expansion: weight each device by the inverse of its selection probability. This requires knowing the panel design and penetration, which commercial providers rarely publish.
- Calibration: adjust the sample so that its marginal totals match known ground truth (traffic counts, transit entries, store door counts, population). This is the most defensible approach for a first-party deployment.
- Model-based correction: train a model to predict ground truth from sample features (device counts, dwell time, POI attributes, weather, time). Placer, SafeGraph, and Unacast all do some version of this [Placer-Data; SafeGraph-Bias; Unacast-Method].
The trap is that model-based correction can hide bias rather than remove it. If the panel under-represents a neighbourhood, a model trained on the same panel will learn the wrong baseline. Calibration against independent counts is essential.
6.3 Forecasting
A demand forecast typically combines:
- Historical baseline: same hour/day-of-week/season, robust to outliers.
- Live signal: current probes, lagged and smoothed.
- Context: weather, holidays, events, school terms, roadworks.
- Model: gradient boosting, temporal fusion, graph neural networks, or a simple seasonal model with prediction intervals.
Google's ETA system is a production example of graph-based spatio-temporal forecasting [Google-ETA-GNN; DeepMind-2020]. TomTom and HERE offer predictive traffic products with historical and real-time inputs [TomTom-Prediction; HERE-RTT]. For footfall, Placer and Unacast produce visit estimates and trends, not necessarily forward forecasts [Placer-Data; Unacast-Method].
Two rules for our paper:
- Report uncertainty. A point forecast without an interval is not decision-grade. Use quantile or conformal intervals, and validate coverage.
- Separate nowcast from forecast. A 15-minute "forecast" that is mostly the current observation is a nowcast. Say so, especially if the data feed is replay or batch.
6.4 Failure modes
| Failure mode | What happens | Mitigation |
|---|---|---|
| Penetration bias | Probe share varies by road, hour, device, and app; counts are not population counts | Estimate penetration where possible; calibrate to counts; publish sample size |
| Panel / app bias | Panel over-represents certain ages, incomes, platforms, or geographies | Compare to ground truth; weight; audit by stratum; disclose |
| Urban canyon | GPS multipath and blocked sky view degrade accuracy; tall buildings and parking garages make the blue dot wrong [Google-LocationAccuracy] | Fuse Wi-Fi/cell; filter outliers; use map matching; report accuracy |
| Indoor coverage | GPS is weak or unavailable indoors; Wi-Fi/cell may be coarse [Google-LocationAccuracy] | Use indoor anchors/Wi-Fi; treat indoor data separately; do not overclaim |
| Device multiplicity | One person carries multiple devices, or one device is shared | Report devices, not people; where possible, de-duplicate at the event level without persistent IDs |
| Road coverage gaps | Minor roads, rural areas, and new roads have few probes | Fall back to historical/functional-class models; mark model-filled segments |
| Temporal drift | App updates, privacy settings, pandemics, construction change penetration | Monitor sample size and bias over time; retrain; change-point detection |
| Privacy suppression | k-threshold/DP removes small cells, which are often the most interesting | Report suppression rate; use complementary suppression; do not silently impute zeros |
7. Implications for our architecture
7.1 Design principles
- Opt-in, default off, granular. Separate consent for probe measurement, product analytics, and research. Withdrawal stops future collection and deletes raw data.
- Minimise at the source. The device should map-match and bin locally; raw GPS/Wi-Fi/cell observations should not leave the device unless necessary.
- No raw trajectories, no persistent IDs. Use rotating pseudonyms or no identifier at all. Never store a per-device trajectory that can be joined across days.
- Local DP before transmission. Add calibrated noise on the device, with a stated ε and a privacy budget. Do not rely on the server to anonymize after receiving raw data.
- Secure aggregation. Use a protocol such as Prio or DAP so that the server sees only aggregate sums, not individual contributions [Prio-2017; DAP-2026]. This protects against a curious server and complements DP.
- k-threshold and suppression. Publish a cell only if at least k distinct contributors are present; suppress or coarsen small cells; exclude sensitive sites (residential, medical, places of worship, schools, shelters).
- Retention limits. Raw data on device only; aggregates retained for a defined period; no cross-day join keys; document backups.
- Transparency. Publish the schema, spatial/temporal bins, k, ε, retention, coverage, and bias audits. If we cannot publish a parameter, we should not claim the guarantee.
7.2 Reference architecture
┌──────────────────────────── Device ────────────────────────────┐
│ GPS / Wi-Fi / cell / IMU │
│ │ │
│ ▼ │
│ map-match to road segment or H3 cell │
│ │ │
│ ▼ │
│ local bin: 15 min (traffic) / 1 h (footfall) │
│ │ │
│ ▼ │
│ local DP: calibrated noise; clip contribution; budget │
│ │ │
│ ▼ │
│ k-threshold check: only send if ≥ k contributors in the bin │
│ │ │
│ ▼ │
│ secret-share / encrypt (Prio, DAP, or 2-of-3 MPC) │
└────────────────────────────┬───────────────────────────────────┘
│ no raw points, no persistent ID
▼
┌──────────────────── Aggregation servers ───────────────────────┐
│ compute sums / averages over shares; never see individual data │
│ tolerate dropouts; verify client contributions │
└────────────────────────────┬───────────────────────────────────┘
│ aggregate only
▼
┌──────────────────── Analytics / product ───────────────────────┐
│ heatmaps, segment speeds, demand indices, forecasts │
│ publish uncertainty, sample counts, suppression rate │
└────────────────────────────────────────────────────────────────┘
This is a privacy-enhanced architecture, not a magical one. It still leaks aggregate information; DP and k-threshold bound that leakage. It still has sampling bias; calibration addresses that. It is auditable in a way that "we anonymize it" is not.
7.3 Concrete defaults
| Parameter | Suggested default | Rationale |
|---|---|---|
| Consent | Off by default; separate toggles for measurement/research/ads | Matches Google's Timeline default and GDPR expectations; avoids consent bundling |
| Spatial bin (traffic) | Road segment + direction | Traffic is a link-level phenomenon; hexagons blur parallel roads |
| Spatial bin (footfall) | H3 resolution 8 (~0.7 km²) or 9 (~0.1 km²) | H3 gives stable hierarchical aggregation [H3]; choose based on venue density |
| Temporal bin | 15 min traffic; 1 h footfall | Balances k-threshold feasibility and operational usefulness |
| k-threshold | 20–50 distinct contributors per cell/time bin | Placer uses 50 [Placer-Data]; Google's search uses an area of at least 3 km² or 1,000 people for a different purpose [Google-LocationData]. Start at 50 for public release, 20 for internal ops |
| Local DP ε | 1–3 per release, with a daily/weekly budget | A starting range to be justified in the DPIA; must be documented and tested. Do not claim DP without stating ε |
| Secure aggregation | 2-of-3 servers or Prio/DAP | Tolerates one honest server; protects against curious aggregator [Prio-2017; DAP-2026] |
| Raw retention | On device only; delete after upload/binning | Minimises breach impact and re-identification risk |
| Aggregate retention | 90 days hot, 1 year cold; no cross-day keys | Enough for seasonal analysis; limits long-term linkage |
| Sensitive sites | Exclude residential, medical, places of worship, schools, shelters, and other sensitive categories | Extends Google's residential exclusion and Placer/SafeGraph sensitive-place policies [Google-BusyAreas; Placer-Trust; SafeGraph-Warren] |
| Transparency | Publish schema, k, ε, retention, coverage, suppression rate, bias audit | Makes claims verifiable; supports peer review |
7.4 Sensing-layer comparison for the paper
| Layer | What it measures | Identifiers | Accuracy for traffic | Coverage | Consent | Privacy risk | Verdict |
|---|---|---|---|---|---|---|---|
| Passive Bluetooth/Wi-Fi MAC sniffing | Presence of discoverable devices; dwell time; rough counts | Device MAC (often randomized) | Poor for speed; count bias | Local receiver coverage; pedestrian/vehicle mix | None (passive) | High: identifiers that may be persistent or linkable when randomization is absent; no consent | Avoid as a primary sensing layer |
| Multi-receiver RSSI / AoA multilateration | Position from signal strength/angle | Device MAC or ephemeral ID | Can be good indoors; multipath outdoors | Requires dense receiver infrastructure | Usually none or venue notice | High if IDs retained; medium if ephemeral and on-device | Use only with ephemeral IDs and aggregation; not a traffic-speed solution |
| Optical / thermal counters | Counts, direction, sometimes dwell | None | N/A (not speed) | Count-line specific | Notice; no personal data if no images | Low if images are not retained | Good ground truth for calibration |
| First-party app probe | Speed, travel time, footfall, dwell | Rotating pseudonym or none | Good where penetration is sufficient; map-matched | App users; coverage depends on adoption | Opt-in | Medium; can be strong with local DP + secure aggregation | Best fit for our proposal |
| Telco mobility | Population presence/flows/O-D | Subscriber ID (internal), aggregated output | Coarse; not lane-level | Population-scale; rural gaps | Legal basis/consent varies | Medium–high; aggregation essential | Good for regional O-D, not street traffic |
| Transit smart card | Transit entries/exits/O-D | Card ID (pseudonymized) | N/A for road traffic | Transit network only | Fare terms/privacy policy | Medium; journey reconstruction risk | Good for transit O-D |
| Manual counts | Counts, classification | None | N/A | Sparse, expensive | Notice | Low | Ground truth, not a live layer |
The table makes the paper's argument concrete: MAC sniffing is typically the weakest option for both reliability and privacy. It measures presence, not movement; it depends on discoverable, often randomized identifiers; it lacks consent; and it cannot produce a defensible speed estimate without a dense, calibrated receiver network. Aggregated first-party probe data is not perfect, but it measures the quantity we actually care about (movement) and can be designed with formal privacy controls.
7.5 What to claim and what not to claim
| Claim | Safe wording | Why |
|---|---|---|
| "We measure traffic" | "We estimate segment speeds/travel times from opt-in probe data, calibrated against ground counts" | Probe data is a sample; "measure" overstates certainty |
| "We count people" | "We estimate devices; penetration is unknown" | Multiple devices per person; shared devices; panel bias |
| "Our data is anonymous" | "Our data is aggregated with a k-threshold and local DP at ε = ..." | "Anonymous" is undefined and often false |
| "We use triangulation" | "We use map-matched probe data" or "multi-receiver multilateration" | Triangulation is angle-based; probe data is not a positioning method |
| "Google triangulates phones" | "Google fuses device-side location and aggregates navigation probes" | The public model does not support the stronger claim |
| "Google uses DP for traffic" | "Google claims DP for busyness; no DP claim for traffic was found" | Distinguishes verified scope |
| "MAC sniffing counts crowds" | "MAC sniffing is a biased presence proxy; randomization breaks uniqueness" | Defensible critique |
| "Our forecast is accurate" | "Our forecast has MAE/WAPE ... on a held-out period, with limitations" | Numbers must match measured results |
7.6 Evaluation plan
- Ground truth. Door counters, loop detectors, transit gate counts, or manual counts. Without ground truth, a probe-based estimate is only a proxy.
- Accuracy metrics. MAE, WAPE, MAPE (with care around zeros), R², and prediction-interval coverage. Report per sensor/segment, not just aggregate.
- Bias audit. Compare errors by neighbourhood, income proxy, device type, time of day, and day of week. Report the worst stratum, not only the mean.
- Privacy audit. State k, ε, retention, and suppression rate. Run a membership-inference or re-identification test on the released aggregates. If we cannot run it, say so.
- Ablation. Compare raw vs aggregated vs local-DP vs secure-aggregation pipelines on the same data. Report the utility cost of each privacy layer.
- Robustness. Missing data, low penetration, urban canyon, indoor venues, holidays, and weather.
- Operational metrics. Latency, update cadence, coverage, and failure modes (e.g., fallback to historical baseline).
7.7 Threats to validity
- Device ≠ person. All device-based counts inherit this.
- Penetration unknown. We can estimate it only with calibration; otherwise we report an index, not a population count.
- Panel skew. Our opt-in sample may not represent the venue's visitors.
- Map-matching errors. Parallel roads, tunnels, and parking lots can misassign probes.
- Privacy suppression. k-threshold and DP remove small cells; the remaining map is biased toward busy areas.
- Company claims. Google/HERE/TomTom/INRIX/Placer/SafeGraph/Unacast numbers are not independently verified; use them as context, not ground truth.
8. Key papers and standards
8.1 Probe data and traffic estimation
- Google Maps ETA with graph neural networks. Derrow-Pinion et al., ETA Prediction with Graph Neural Networks in Google Maps, CIKM 2021. Production GNN over road supersegments; strong applied reference [Google-ETA-GNN].
- DeepMind traffic prediction. Lange & Perez, Traffic prediction with advanced Graph Neural Networks, 2020. Accessible description of supersegments and multi-loss training [DeepMind-2020].
- Mobile Century. Herrera et al., Evaluation of traffic data obtained via GPS-enabled mobile phones: The Mobile Century field experiment, Transportation Research Part C, 2009. 100 vehicles; suggests 2–3% phone penetration can be enough for velocity estimation [Herrera-2009].
- Probe penetration and flow estimation. Seo et al., Estimation of flow and density using probe vehicles with spacing measurement equipment, Transportation Research Part C, 2015. 3.5% penetration for 5-minute flow; 0.2% for hourly flow [Seo-2015].
- Sampling protocol bias. Jenelius & Koutsopoulos, Probe vehicle data sampled by time or space, Transportation Research Part B, 2015. Sampling protocol affects travel-time bias [Jenelius-2015].
- Loop vs probe trade-offs. Mazaré et al., Trade-offs between inductive loops and GPS probe vehicles for travel time estimation: A Mobile Century case study, TRB 2012. Penetration 1.5–3% at 275 veh/h; diminishing returns [Mazare-2011].
- Empirical penetration distribution. Luo et al., Empirical distribution of probe vehicle penetration rates and implications for traffic state estimation, Transportation Research Part C, 2026. Penetration is non-uniform; a reliable penetration estimate matters more than assuming uniformity [Luo-2026].
- Virtual trip lines. Hoh et al., Virtual Trip Lines for Distributed Privacy-Preserving Traffic Monitoring, MobiSys 2008; and Enhancing Privacy and Accuracy in Probe Vehicle-Based Traffic Monitoring via Virtual Trip Lines, IEEE TMC 2012. Spatial triggers reduce raw trajectory exposure and support distributed aggregation [Hoh-2008; Hoh-2012].
- Differential privacy for traffic flow. Le Ny et al., Real-time privacy-preserving model-based estimation of traffic flows, ICCPS 2014. DP traffic-state estimation with a macroscopic model (useful for the privacy-utility trade-off) [LeNy-2014].
8.2 Privacy foundations
- k-anonymity. Sweeney, k-anonymity: A Model for Protecting Privacy, 2002 [Sweeney-2002].
- Differential privacy. Dwork, Differential Privacy, ICALP 2006 [Dwork-2006]; Dwork & Roth, The Algorithmic Foundations of Differential Privacy, 2014 [Dwork-Roth-2014].
- Geo-indistinguishability. Andrés et al., Geo-Indistinguishability: Differential Privacy for Location-Based Systems, CCS 2013. Formal location privacy with a radius-dependent guarantee [Andres-2013].
- Local DP in practice. Erlingsson et al., RAPPOR: Randomized Aggregatable Privacy-Preserving Ordinal Response, CCS 2014. Google's client-side randomized response deployment [RAPPOR-2014].
- Secure aggregation. Bonawitz et al., Practical Secure Aggregation for Privacy-Preserving Machine Learning, CCS 2017 [Bonawitz-2017].
- Prio. Corrigan-Gibbs & Boneh, Prio: Private, Robust, and Scalable Computation of Aggregate Statistics, NSDI 2017 [Prio-2017].
- DAP. IETF PPM Working Group, Distributed Aggregation Protocol for Privacy Preserving Measurement, draft-ietf-ppm-dap-20, 2026 [DAP-2026].
- Aggregate location privacy. Pyrgelis et al., Knock Knock, Who's There? Membership Inference on Aggregate Location Data, NDSS 2018 [Pyrgelis-2018].
- Mobility uniqueness. de Montjoye et al., Unique in the Crowd: The privacy bounds of human mobility, Scientific Reports, 2013 [deMontjoye-2013].
8.3 Standards and frameworks
- NIST SP 800-188. De-Identifying Government Datasets, 2023. Practical de-identification and risk assessment [NIST-800-188].
- ISO/IEC 20889:2018. Privacy enhancing data de-identification terminology and classification of techniques [ISO-20889].
- DATEX II / EN 16157. European standard for traffic and travel information exchange; measured and elaborated data publications [DATEX-II].
- ISO 14819-1:2021. Traffic and travel information messages via traffic message coding; RDS-TMC/ALERT-C [ISO-14819].
- H3. Open-source hexagonal geospatial indexing system, suitable for hierarchical aggregation [H3].
8.4 Data-product documentation
- Google Maps traffic/ETA, Timeline, busyness, anonymization, retention [Google-Maps101; Google-NavData; Google-BusyAreas; Google-Timeline2023; Google-Anonymization; Google-Retention].
- HERE, TomTom, INRIX, Waze product and methodology pages [HERE-RTT; HERE-Probe; TomTom-Prediction; TomTom-API; INRIX; Waze-Cities].
- Placer, SafeGraph, Unacast methodology and privacy pages [Placer-Data; Placer-Trust; SafeGraph-Privacy; SafeGraph-Warren; SafeGraph-Bias; Unacast-Method; Unacast-Legal].
- Telco, transit, and fleet examples [Telefonica; Orange-Flux; Orange-Cambridge; Vodafone-Cambridge; TfL-OpenData; TfL-Access; Geotab-FAQ; Geotab-Privacy].
9. Citations
All URLs accessed 2026-11-14 unless otherwise stated. Company pages may change; the title and access date are part of the citation.
[Google-Maps101] Johann Lau, "Google Maps 101: How AI helps predict traffic and determine routes," Google, 3 September 2020. https://blog.google/products/maps/google-maps-101-how-ai-helps-predict-traffic-and-determine-routes/
[Google-NavData] Google Maps Help, "How navigation data makes Maps better for everyone." https://support.google.com/maps/answer/10565726
[Google-TrafficLegend] Google Maps Help, "Use layers to find places, traffic, terrain, biking & transit." https://support.google.com/maps/answer/3092439
[Google-PrivacyPolicy] Google, "Privacy Policy." https://policies.google.com/privacy
[Google-LocationData] Google Privacy & Terms, "How Google uses location information." https://policies.google.com/technologies/location-data
[Google-LocationAccuracy] Google Maps Help, "Find & improve your location's accuracy in Google Maps." https://support.google.com/maps/answer/2839911
[Google-BusyAreas] Google Maps Help, "Get information about busy areas from Google Maps." https://support.google.com/maps/answer/11323117
[Google-TimelinePrivacy] Google Maps Help, "How Google keeps your Timeline data private." https://support.google.com/maps/answer/10077010
[Google-TimelineManage] Google Account Help, "Manage your Timeline data." https://support.google.com/accounts/answer/14200149
[Google-Timeline2023] Marlo McGriff, "Updates to Location History and new controls coming soon to Maps," Google, 12 December 2023. https://blog.google/products/maps/updates-to-location-history-and-new-controls-coming-soon-to-maps/
[Google-Incognito] Google Maps Help, "Use Google Maps in Incognito mode." https://support.google.com/maps/answer/9430563
[Google-LocationToggle] Google Maps Help, "Turn location data collection on or off." https://support.google.com/maps/answer/2839958
[Google-Anonymization] Google Privacy & Terms, "How Google anonymizes data." https://policies.google.com/technologies/anonymization
[Google-Retention] Google Privacy & Terms, "How Google retains data we collect." https://policies.google.com/technologies/retention
[Google-RMI] Google for Developers, "Roads Management Insights: Data foundation." https://developers.google.com/maps/documentation/roads-management-insights/data
[Google-GeolocationAPI] Google for Developers, "Geolocation API Overview." https://developers.google.com/maps/documentation/geolocation/overview
[Google-2009] Google Official Blog, "The bright side of sitting in traffic: Crowdsourcing road congestion data," 25 August 2009. https://googleblog.blogspot.com/2009/08/bright-side-of-sitting-in-traffic.html
[Google-AU-2009] Andrew Foster, "Get there faster with traffic information on Google Maps," Google Australia Blog, 15 September 2009. https://australia.googleblog.com/2009/09/get-there-faster-with-traffic.html
Positioning and terminology
[GPSgov-Trilateration] GPS.gov, "Trilateration." https://www.gps.gov/trilateration
[NOAA-Geodesy] NOAA National Geodetic Survey, "Geodesy for the Layman" (triangulation and trilateration definitions). https://geodesy.noaa.gov/PUBS_LIB/Geodesy4Layman/TR80003A.HTM
[Android-RTT] Android Developers, "Wi-Fi location: ranging with RTT." https://developer.android.com/develop/connectivity/wifi/wifi-rtt
Alternatives
[HERE-RTT] HERE, "Real-Time Traffic," HERE Traffic API v7 documentation. https://docs.here.com/traffic-api/docs/real-time-traffic-here-traffic-api-v7-concepts
[HERE-OnePager] HERE, "HERE Real-Time Traffic" product one-pager, 2026. https://assets.here.com/x7rx8ayph7ee/e13d7d3e-c1f8-4562-876c-e4910c7001a8/9e056bbd245dab0a0587e5992b55f4d5/HERE_Traffic_product_one_pager.pdf
[HERE-Probe] HERE, "GPS probe data for Traffic Analytics." https://docs.here.com/traffic-analytics/docs/gps-probe-data
[TomTom-Prediction] Joao Laranjeira, "What is Traffic Prediction and How Does it Work?" TomTom Newsroom, 24 January 2020. https://www.tomtom.com/newsroom/behind-the-map/road-traffic-prediction
[TomTom-Stats] TomTom, "TomTom Traffic Stats explained." https://www.tomtom.com/newsroom/product-focus/tomtom-traffic-stats-explained
[TomTom-API] TomTom, "Traffic Stats API: Introduction." https://docs.tomtom.com/traffic-stats/documentation/product-information/introduction
[INRIX] INRIX, "INRIX AI Traffic." https://inrix.com/products/ai-traffic/
[Waze-Cities] Waze, "Waze for Cities." https://www.waze.com/wazeforcities
[IowaDOT-2020] "DOT Traveler Information Website Crowdsourcing Practices," Project 14.4, 2020 (Iowa DOT comparison of Google and Waze traffic layers). https://nwpassage.info/wp-content/uploads/2024/04/14-4-crowdsourcing-practices.pdf
[Apple-Privacy] Apple, "Apple Privacy Features: Maps." https://www.apple.com/privacy/features/
[Placer-Data] Placer.ai, "Our Data." https://www.placer.ai/resources/our-data
[Placer-Trust] Placer.ai, "Trust Center." https://www.placer.ai/company/trust-center
[SafeGraph-Privacy] SafeGraph, "Privacy Policy & Data Practices," last updated 2 December 2025. https://www.safegraph.com/privacy-policy
[SafeGraph-Warren] SafeGraph, response to Senator Elizabeth Warren, 23 June 2022. https://www.warren.senate.gov/wp-content/uploads/media/doc/SafeGraph%20Response%20%20to%20May%2017%20Letter_Redacted.pdf
[SafeGraph-Bias] Wang et al., "Understanding the bias of mobile location data across spatial scales and over time: A comprehensive analysis of SafeGraph data in the United States," PLOS One, 2024. https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0294430
[Unacast-Method] Unacast, "Our Methodology." https://www.unacast.com/methodology
[Unacast-Legal] Unacast, "Legal." https://docs.unacast.com/legal
[Telefonica] Telefónica Germany, "Mobility Insights: Anonymization process." https://www.telefonica.de/anonymization_process
[Orange-Flux] Orange Business, "Flux Vision." https://www.orange-business.com/en/solutions/data-intelligence-iot/flux-vision
[Orange-Cambridge] "The value of network data confirmed by the Covid-19 epidemic and its expanded usages," Data & Policy, 2022. https://www.cambridge.org/core/journals/data-and-policy/article/value-of-network-data-confirmed-by-the-covid19-epidemic-and-its-expanded-usages/4FD137D39AF08DB624981FDA322CABB0
[Vodafone-Cambridge] "Using Vodafone mobile phone network data to provide insights into citizens mobility in Italy during the Coronavirus outbreak," Data & Policy, 2021. https://www.cambridge.org/core/journals/data-and-policy/article/using-vodafone-mobile-phone-network-data-to-provide-insights-into-citizens-mobility-in-italy-during-the-coronavirus-outbreak/100EA32E60F4A6E64AE96515B2A80EA4
[TfL-OpenData] Transport for London, "Our open data." https://tfl.gov.uk/info-for-open-data-users/our-open-data
[TfL-Access] Transport for London, "Access your data." https://lul-trackernet.tfl.gov.uk/corporate/privacy-and-cookies/access-your-data
[TfL-Oyster] Transport for London, "Oyster card" privacy page. https://lul-trackernet.tfl.gov.uk/corporate/privacy-and-cookies/oyster-card
[Geotab-FAQ] Altitude by Geotab, "Geolocation Data Privacy FAQ," 14 October 2025. https://altitude.geotab.com/news-and-releases/geolocation-data-privacy-faq
[Geotab-Privacy] Geotab, "Our Privacy Commitment." https://www.geotab.com/about/privacy
Probe data and privacy papers
[LeNy-2014] J. Le Ny, A. Touati, and G. J. Pappas, "Real-time privacy-preserving model-based estimation of traffic flows," ACM/IEEE ICCPS 2014. DOI: 10.1109/ICCPS.2014.6843714.
[Hoh-2008] B. Hoh, M. Gruteser, R. Herring, J. Ban, D. Work, J. C. Herrera, A. M. Bayen, M. Annavaram, and Q. Jacobson, "Virtual Trip Lines for Distributed Privacy-Preserving Traffic Monitoring," MobiSys 2008. DOI: 10.1145/1378600.1378604.
[Hoh-2012] B. Hoh, T. Iwuchukwu, Q. Jacobson, D. Work, A. M. Bayen, R. Herring, J.-C. Herrera, M. Gruteser, M. Annavaram, and J. Ban, "Enhancing Privacy and Accuracy in Probe Vehicle-Based Traffic Monitoring via Virtual Trip Lines," IEEE Transactions on Mobile Computing, 11(5), 849–864, 2012. DOI: 10.1109/TMC.2011.116.
[Herrera-2009] J. C. Herrera, D. B. Work, R. Herring, X. Ban, Q. Jacobson, and A. M. Bayen, "Evaluation of traffic data obtained via GPS-enabled mobile phones: The Mobile Century field experiment," Transportation Research Part C, 2009.
[Seo-2015] T. Seo, T. Kusakabe, and Y. Asakura, "Estimation of flow and density using probe vehicles with spacing measurement equipment," Transportation Research Part C, 2015.
[Jenelius-2015] E. Jenelius and H. N. Koutsopoulos, "Probe vehicle data sampled by time or space: Consistent travel time allocation and estimation," Transportation Research Part B, 2015.
[Mazare-2011] P.-E. Mazaré, O.-P. Tossavainen, A. M. Bayen, and D. B. Work, "Trade-offs between inductive loops and GPS probe vehicles for travel time estimation: A Mobile Century case study," Transportation Research Board, 2012.
[Luo-2026] Y. Luo, J. Balsa-Barreiro, H. Yi, and M. Menéndez, "Empirical distribution of probe vehicle penetration rates and implications for traffic state estimation," Transportation Research Part C, 189, 105729, 2026. DOI: 10.1016/j.trc.2026.105729.
[Google-ETA-GNN] A. Derrow-Pinion et al., "ETA Prediction with Graph Neural Networks in Google Maps," CIKM 2021, arXiv:2108.11482. DOI: 10.1145/3459637.3481916.
[DeepMind-2020] O. Lange and L. Perez, "Traffic prediction with advanced Graph Neural Networks," Google DeepMind, 3 September 2020. https://deepmind.google/discover/blog/traffic-prediction-with-advanced-graph-neural-networks/
[Sweeney-2002] L. Sweeney, "k-anonymity: A Model for Protecting Privacy," International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems, 10(5), 557–570, 2002. DOI: 10.1142/S0218488502001648.
[Dwork-2006] C. Dwork, "Differential Privacy," ICALP 2006, LNCS 4052, pp. 1–12. DOI: 10.1007/11787006_1.
[Dwork-Roth-2014] C. Dwork and A. Roth, "The Algorithmic Foundations of Differential Privacy," Foundations and Trends in Theoretical Computer Science, 2014. https://www.cis.upenn.edu/~aaroth/Papers/privacybook.pdf
[Andres-2013] M. E. Andrés, N. E. Bordenabe, K. Chatzikokolakis, and C. Palamidessi, "Geo-Indistinguishability: Differential Privacy for Location-Based Systems," CCS 2013. DOI: 10.1145/2508859.2516735.
[RAPPOR-2014] Ú. Erlingsson, V. Pihur, and A. Korolova, "RAPPOR: Randomized Aggregatable Privacy-Preserving Ordinal Response," CCS 2014. DOI: 10.1145/2660267.2660348.
[Bonawitz-2017] K. Bonawitz et al., "Practical Secure Aggregation for Privacy-Preserving Machine Learning," CCS 2017, pp. 1175–1191. DOI: 10.1145/3133956.3133982.
[Prio-2017] H. Corrigan-Gibbs and D. Boneh, "Prio: Private, Robust, and Scalable Computation of Aggregate Statistics," NSDI 2017. https://www.usenix.org/conference/nsdi17/technical-sessions/presentation/corrigan-gibbs
[DAP-2026] IETF Privacy Preserving Measurement Working Group, "Distributed Aggregation Protocol for Privacy Preserving Measurement," draft-ietf-ppm-dap-20, October 2026. https://datatracker.ietf.org/doc/draft-ietf-ppm-dap/
[Pyrgelis-2018] A. Pyrgelis, C. Troncoso, and E. De Cristofaro, "Knock Knock, Who's There? Membership Inference on Aggregate Location Data," NDSS 2018. https://www.ndss-symposium.org/wp-content/uploads/2018/02/ndss2018_05B-5_Pyrgelis_paper.pdf
[deMontjoye-2013] Y.-A. de Montjoye, C. A. Hidalgo, M. Verleysen, and V. D. Blondel, "Unique in the Crowd: The privacy bounds of human mobility," Scientific Reports, 3, 1376, 2013. DOI: 10.1038/srep01376.
[NIST-800-188] S. Garfinkel, J. Near, A. Dajani, P. Singer, and B. Guttman, "De-Identifying Government Datasets," NIST Special Publication 800-188, 2023. DOI: 10.6028/NIST.SP.800-188.
[ISO-20889] ISO/IEC 20889:2018, "Privacy enhancing data de-identification terminology and classification of techniques." https://www.iso.org/standard/69373.html
[DATEX-II] DATEX II, "DATEX II Documentation Portal" (CEN/EN 16157 series). https://docs.datex2.eu/
[ISO-14819] ISO 14819-1:2021, "Intelligent transport systems — Traffic and travel information messages via traffic message coding — Part 1: Coding protocol for Radio Data System-Traffic Message Channel (RDS-TMC) using ALERT-C." https://www.iso.org/standard/76781.html
[H3] Uber, "H3: A hexagonal hierarchical geospatial indexing system." https://h3geo.org/docs/
Appendix A. One-paragraph summary for the paper's related-work section
Google Maps' traffic and ETA features are built on aggregated, map-matched probe data, not, as publicly described, on cell-tower triangulation. Google fuses GPS, Wi-Fi, cellular, and sensor signals on the device; while users navigate, it collects GPS, route, and sensor data; it aggregates those observations into road-segment speeds and combines them with historical traffic patterns and graph-neural-network ETA models [Google-Maps101; Google-NavData; Google-ETA-GNN]. Google's public privacy claims are layered: aggregation and anonymization for traffic, differential privacy for busyness/popular times, and on-device Timeline storage with a three-month default auto-delete [Google-BusyAreas; Google-Timeline2023]. These are meaningful protections, but they are not a substitute for a formal, auditable guarantee: Google does not publish penetration, aggregation thresholds, DP epsilon, traffic-layer coverage, or an independent audit, and aggregate location data remains vulnerable to membership inference [Pyrgelis-2018]. The lesson for a privacy-first sensing layer is to make those parameters explicit: opt-in consent, on-device aggregation, local DP, secure aggregation, a k-threshold, no raw trajectories, and bounded retention.
Appendix B. Quick fact table
| Fact | Value | Source |
|---|---|---|
| Google Maps daily distance | Over 1 billion km/day | [Google-Maps101] |
| Google Maps country coverage | 220+ countries and territories | [Google-Maps101] |
| Google ETA accuracy claim | "Over 97% of trips" | [Google-Maps101] |
| DeepMind ETA improvement claim | Up to 50% in some cities | [Google-Maps101; DeepMind-2020] |
| Google GPS accuracy (user-facing) | About 20 m | [Google-LocationAccuracy] |
| Google cell accuracy (user-facing) | Up to a few thousand metres | [Google-LocationAccuracy] |
| Google search location threshold | 3 km² or 1,000 people | [Google-LocationData] |
| Timeline default auto-delete | 3 months (2023 update) | [Google-Timeline2023] |
| Google deletion time | ~2 months; backups up to 6 months | [Google-Retention] |
| HERE traffic coverage | 75 countries; flow every minute | [HERE-RTT; HERE-OnePager] |
| TomTom probe base | ~600 million probes | [TomTom-Prediction] |
| Waze driver base | ~200 million drivers | [Waze-Cities] |
| Placer k-anonymity | 50 devices | [Placer-Data; Placer-Trust] |
| TfL Oyster retention | 8–9 weeks | [TfL-Access] |
| Mobile Century penetration | 2–3% enough for velocity | [Herrera-2009] |
| Seo et al. penetration | 3.5% for 5-min flow; 0.2% hourly | [Seo-2015] |